Fine-Tuning and Efficient VGG16 Transfer Learning Fault Diagnosis Method for Rolling Bearing

Abstract

Nowadays, neural network become popular in modeling. However, the model training needs a lot of data, long training time and high hardware conditions. It is inefficient for ordinary computing devices to be used in training models. In this paper, VGG16 model was modified to fit ten labels and used as feature extractor. The default image size of model was 224 × 224 pixels. Then the images were reduced into low resolution as 112 × 112, 75 × 75, 56 × 56, 45 × 45, 32 × 32 pixels, which were 1/2, 1/3, 1/4, 1/5 of default side length and the minimum size. Next these images were sent to model for training. The training results illustrated that the images of 112 × 112, 75 × 75, 56 × 56 groups can still be adequate for modified VGG16 to classified and achieve high accuracy and meanwhile significantly reduce the training time. However, when the size dropped to 45 × 45, 32 × 32, overfitting appears and the training accuracy significantly dropped. Thus, it is recommended that set a target accuracy first and begin training from a small size. If the accuracy was not high enough, enlarge the size and train again.